Air Pollution Exposure and COVID-19 - IZA DP No. 13367 JUNE 2020 Matthew A. Cole Ceren Ozgen Eric Strobl - Institute of Labor ...
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
DISCUSSION PAPER SERIES IZA DP No. 13367 Air Pollution Exposure and COVID-19 Matthew A. Cole Ceren Ozgen Eric Strobl JUNE 2020
DISCUSSION PAPER SERIES
IZA DP No. 13367
Air Pollution Exposure and COVID-19
Matthew A. Cole
University of Birmingham
Ceren Ozgen
University of Birmingham and IZA
Eric Strobl
University of Bern
JUNE 2020
Any opinions expressed in this paper are those of the author(s) and not those of IZA. Research published in this series may
include views on policy, but IZA takes no institutional policy positions. The IZA research network is committed to the IZA
Guiding Principles of Research Integrity.
The IZA Institute of Labor Economics is an independent economic research institute that conducts research in labor economics
and offers evidence-based policy advice on labor market issues. Supported by the Deutsche Post Foundation, IZA runs the
world’s largest network of economists, whose research aims to provide answers to the global labor market challenges of our
time. Our key objective is to build bridges between academic research, policymakers and society.
IZA Discussion Papers often represent preliminary work and are circulated to encourage discussion. Citation of such a paper
should account for its provisional character. A revised version may be available directly from the author.
ISSN: 2365-9793
IZA – Institute of Labor Economics
Schaumburg-Lippe-Straße 5–9 Phone: +49-228-3894-0
53113 Bonn, Germany Email: publications@iza.org www.iza.org
IZA DP No. 13367 JUNE 2020
ABSTRACT
Air Pollution Exposure and COVID-19*
In light of the existing preliminary evidence of a link between Covid-19 and poor air quality,
which is largely based upon correlations, we estimate the relationship between long term
air pollution exposure and Covid-19 in 355 municipalities in the Netherlands. Using detailed
secondary and administrative data we find compelling evidence of a positive relationship
between air pollution, and particularly PM2.5 concentrations, and Covid-19 cases, hospital
admissions and deaths. This relationship persists after controlling for a wide range of
explanatory variables. Our results indicate that a 1 μ/m3 increase in PM2.5 concentrations
is associated with 9.4 more Covid-19 cases, 3.0 more hospital admissions, and 2.3 more
deaths. The relationship between Covid-19 and air pollution withstands a number of
sensitivity and robustness exercises including instrumenting pollution to mitigate potential
endogeneity and modelling spatial spillovers using spatial econometric techniques.
JEL Classification: I21, I23, Q53
Keywords: COVID-19, air pollution, Netherlands, spatial spillovers
Corresponding author:
Matthew A. Cole
University of Birmingham
Department of Economics
University House
Edgbaston, B15 2TT
United Kingdom
E-mail: m.a.cole@bham.ac.uk
* We would like to thank Guus Velders and Hans Koster for the provision of RIVM air pollution data and commuting
time data, respectively.
1 Introduction
The Covid-19 pandemic is causing significant social and economic impacts across large
parts of the world. At the time of writing the number of Covid-19 cases worldwide has
reached 7.2 million, while the death toll has exceeded 400,000.1 Governments and health
care systems are facing the immense challenge of trying to control the spread of the virus
and to prevent hospitals from being overwhelmed as millions of individuals remain subject
to lockdown and face significant economic uncertainty. In order to respond to these un-
precedented challenges it is important for policy makers and health care professionals to
understand which groups of individuals suffer the highest morbidity and mortality risks
from Covid-19 and which factors may exacerbate these risks.
A contributory factor that has been tentatively explored in several recent academic
studies is poor air quality. While some such studies have identified the significant improve-
ments in air quality that have resulted from Covid-19 lockdowns (Cicala et al., 2020; Cole
et al., 2020), others have pointed to a correlation between Covid-19 hotspots and areas with
high levels of pollution concentrations (Travaglio et al., 2020; Conticini et al., 2020). It is
well known that long term exposure to pollutants such as nitrogen dioxide (N O2 ), sulphur
dioxide (SO2 ), and fine particulate matter (P M2.5 ) contributes to cardiovascular disease,
reduces lung function, and causes respiratory illness (Faustini et al., 2014; Ming Han et
al., 2015; Katanoda et al., 2011; Abbey et al., 1999; Weerdt et al., 2020). These pollutants
have been shown to cause a persistent inflammatory response even in the relatively young,
and to increase the risk of infection by viruses that target the respiratory tract (Travaglio
et al., 2020; Conticini et al., 2020). While Covid-19 produces only mild symptoms for most
sufferers, in a minority of cases it results in an excessive inflammatory response causing
Acute Respiratory Distress Syndrome (ARDS) and death. Given the clear overlaps be-
tween the symptoms of Covid-19-induced ARDS and long term exposure to air pollution,
a number of studies have begun to explore the links between the two.
Focusing on the UK, Travaglio et al. (2020), for instance, find evidence of a correlation
between Covid-19 cases and concentrations of nitrogen oxides and ozone, while Ogen (2020)
examines 66 regions across Italy, France, Germany, and Spain and finds similar evidence.
1
As of June 9th 2020. Source https://www.worldometers.info/coronavirus/countries
1
Conticini et al. (2020) focus on Northern Italy and conclude that pollution concentrations
are a likely contributor to the high Covid-19 death rates experienced in that region. A
preliminary analysis of the Netherlands also finds evidence of a link between concentrations
of P M2.5 and Covid-19 cases (Andree, 2020). Finally, Wu et al. (2020) examine US counties
and econometrically estimate the relationship between county-level Covid-19 death rates
and long-term concentrations of fine particulate matter (P M2.5 ), controlling for a wide
range of confounding factors. They conclude that a 1 µg/m3 increase in P M2.5 is associated
with an 8% increase in the Covid-19 death rate.
While the above studies provide useful preliminary evidence, Conticini et al. (2020)
and Ogen (2020) offer only geographical correlations between Covid-19 cases and pollution
exposure, whereas Travaglio et al. (2020) take a similar approach but control only for differ-
ences in population density and do so across only 7 relatively large regions. Establishing a
convincing link between exposure to pollution and Covid-19 cases requires individual-level
data with the ability to control for individual characteristics, such as age and the presence
of underlying health conditions. Since individual-level data on Covid-19 infections is not
available, the next best alternative is to examine a large number of small geographic re-
gions with detailed data on the characteristics of those regions. This allows the researcher
to assess whether any correlation between Covid-19 and pollution exposure still holds once
differences in social deprivation, population density, ethnic composition, and other factors
are controlled for. While Wu et al. (2020) come closest to doing this, US counties are still
relatively large, raising the question of how well such aggregated data can capture the local
variation in confounding effects without being ‘averaged out’.2 Furthermore, by focusing
only on particulate matter, and on Covid-19 deaths, it is not clear whether other pollu-
tants have an effect on Covid-19 deaths once other factors are controlled for or, indeed, on
Covid-19 infections, or hospitalisation rates more generally.
With the above in mind, this paper undertakes a detailed econometric analysis of the
relationship between pollution concentrations and Covid-19 using data for 355 relatively
small Dutch municipalities. More specifically, by using high-resolution air pollution data as
well as combining administrative with secondary data, we estimate the relationship between
2
For instance, Los Angeles County, the largest in the US, had a population of over 10 million in 2019,
with the average county-level population being over 104,000 (US Census Bureau).
2
long term exposure to concentrations of P M2.5 , N O2 , and SO2 and the number of Covid-19
infections, the number of individuals hospitalised with Covid-19, and the number of those
who died as a result of Covid-19. We are able to control for a wide range of potential
confounding effects, including those relating to income, age, underlying health conditions,
education, social deprivation, ethnic composition, workplace characteristics, spatial and
social proximity to potential risk factors and others. Our analysis utilises Covid-19 data
up to 5th June 2020 allowing us to capture almost the entire course of the epidemic and
hence much more fully than the previous studies which have examined data up to only
March or early April. Finally, we undertake a number of robustness exercises including
instrumenting pollution to address possible endogeneity concerns.
The Netherlands provides a useful setting in which to examine the link between Covid-
19 and air pollution. As a relatively small, densely populated nation with an ethnically
diverse, aging population, the country faces a number of potential Covid-19 risk factors. It
also shares an open land border with Germany and Belgium, the latter a country that cur-
rently has the highest number of Covid-19 deaths per capita. The Netherlands additionally
experiences hot spots of poor local air pollution both within urban areas and also, in the
case of P M2.5 , in more rural areas due in part to intensive livestock farming. By early June
2020, the Netherlands had experienced over 6,000 deaths as a result of Covid-19, resulting
in the 7th highest number of Covid-19 deaths per capita.3
Our results provide compelling evidence of a statistically significant positive relation-
ship between air pollution and Covid-19 cases, hospital admissions and deaths. More
specifically, we find that an increase in P M2.5 concentrations of 1 µg/m3 is associated with
an increase in Covid-19 cases of between 9.4 and 15.1, depending on our model. The same
increase in P M2.5 is associated with an increase in Covid-19 hospital admissions of between
2.9 and 4.4, and an increase in Covid-19 deaths of between 2.2 and 3.6.
The remainder of this paper is organized as follows. Section 2 provides background
information on Covid-19 and air pollution in the Netherlands and Section 3 presents our
empirical methodology. Section 4 describes our data, Section 5 reports our results and
Section 6 concludes.
3
As of June 9th 2020. All Dutch Covid-19 data is provided by the National Institute for Public Health
and the Environment (RIVM). The international ranking excludes the principalities of San Marino and
Andorra and stems from: https://www.worldometers.info/coronavirus/countries
3
2 Covid-19 and Air pollution
The first confirmed Covid-19 case in the Netherlands occurred in late February 2020,
and by early June over 46,000 cases had been identified. Daily cases of Covid-19 peaked
at 1,335 on April 10th, while the daily death toll peaked 3 days earlier at 235. Both daily
cases and daily deaths have been declining steadily since. While all 12 of the Netherlands’
provinces have experienced broadly similar trends, levels of Covid-19 cases have differed
significantly across provinces. For instance, in the southern provinces of North Brabant and
South Holland daily cases peaked at over 350, while in other provinces such as Groningen
and Drenthe, both in the north of the country, cases peaked at fewer than 40 per day.
From the outset of the epidemic within the Netherlands the south east has experienced
a disproportionate number of Covid-19 cases. Figure 1 shows the distribution of cases
per capita across the 355 municipalities of the Netherlands up to June 5th 2020. The red
‘hotspots’ in the south east, which are largely within the provinces of North Brabant and
Limburg, demonstrate the relative intensity of cases in those regions. Unusually, these
are relatively rural regions with low population density raising the question of why these
provinces have been so badly affected by Covid-19.4
One potential explanation raised by the Dutch media has been the annual carnival
celebrations held during the last week of February and beginning of March which were
concentrated largely within North Brabant and Limburg. These celebrations attracted
thousands of people from all over the country to street parties and parades, as they do
each year. Numerous Dutch media reports have suggested that these celebrations may at
least partially explain the rapid spread of Covid-19 within these regions and also to other
regions of the Netherlands via carnival participants.5 However, the significant variation
4
Figure 2 in the Appendix shows that both hospital admissions and deaths per capita from Covid-19
have also been disproportionately high in these south eastern regions.
5
A non-exhaustive list includes:
https://www.nrc.nl/nieuws/2020/04/09/vuile-lucht-vergroot-de-kans-om-te-sterven-aan-covid-19-
a3996388
https://nos.nl/nieuwsuur/artikel/2332460-gebieden-met-veel-luchtverontreiniging-zwaarder-getroffen-
door-corona.html
https://www.rtlnieuws.nl/nieuws/nederland/artikel/5094341/corona-update-noorden-minder-
besmettingen
https://nadavos.nl/nieuws/2020-luchtverontreiniging-covid-19
https://www.trouw.nl/binnenland/zorgt-de-veehouderij-voor-meer-coronadoden-in-oost-
brabant b2b5487d/
4
in Covid-19 cases across the municipalities within these two provinces suggests that the
carnival is unlikely to fully explain the density and distribution of these cases.
Another possible explanation that has received less attention in the Netherlands re-
volves around the intensive livestock farming that takes place within North Brabant and
Limburg. These regions house over 63% of the Netherlands’ 12 million pigs and 42% of
its 101 million chickens.6 Such intensive livestock production produces large quantities of
ammonia (N H3 ), which is an important contributor to P M2.5 concentrations. Figure 3
in the Appendix provides a map of 2019 N H3 concentrations and shows that the south
eastern regions have some of the highest concentrations within the Netherlands. Indeed, a
comparison with the map of P M2.5 concentrations in Figure 1 shows a very similar pattern
illustrating how ammonia contributes to P M2.5 .
While the maximum annual average concentration of P M2.5 at municipality level is
12.3 µg/m3 , the equivalent figure for a 1km x 1km gridcell is 23.9µg/m3 , which is close to
the EU air quality standard of 25µg/m3 . This indicates that local concentrations of P M2.5
within the regions of North Brabant and Limburg are close to dangerous levels even when
averaged annually, raising the likelihood that for shorter periods they may exceed safe
levels. Since P M2.5 concentrations show a similar spatial distribution to Covid-19 cases,
also in Figure 1, it’s reasonable to examine this potential link and to see whether the visual
relationship between Covid-19 cases and P M2.5 concentrations withstands controlling for
other potential contributory factors.7
3 Methodology
In order to examine the relationship between Covid-19 and air pollution we begin by
estimating Equation 1 for 355 municipalities:
0 0 0 0 0
Ci = φpollutioni + β1 Di + β2 Pi + β3 Ei + β3 Si + β4 Li + γr + i (1)
6
2019 data from CBS Statline.
7
This article in Trouw, a Dutch national newspaper, expresses the concerns of local residents that
poor air quality in the intensively livestocked regions may be leaving the local population suscepti-
ble to Covid-19, https://www.trouw.nl/binnenland/zorgt-de-veehouderij-voor-meer-coronadoden-in-oost-
brabant b2b5487d/
5
Figure 1: Covid-19 cases per 100,000 people and annual concentrations of
P M2.5 , N O2 and SO2 averaged over the period 2015-19
where C refers to Covid-19 cases, the number of individuals hospitalised by Covid-19,
or the number of deaths from Covid-19 in municipality i as of June 5th 2020. Pollution
refers to annual concentrations of P M2.5 , N O2 , or SO2 , averaged over the period 2015-
0 0 0 0 0
2019. Vectors D , P , E , S and L contain control variables capturing demography, social
6and physical proximity, employment/education, spatial and health variables, respectively,
for the year 2019, as defined in the next section. The term γr denotes province level fixed
effects for each of our 12 provinces, r.
Since our dependent variables take the form of discrete count variables, estimating
Equation 1 using OLS could result in inconsistent, inefficient, and biased estimates (Long,
1997; Hoffman, 2004). The alternative is to use Poisson or negative binomial count models.
Since the error assumption of the Poisson model requires the conditional mean to equal
the conditional variance, a condition that our over-dispersed data fails to meet, we employ
the negative binomial model. This model builds upon the Poisson model by adding a
parameter that allows the conditional variance to exceed the conditional mean.
Importantly our attempt to estimate the relationship between pollution exposure and
Covid-19 cases (as well as hospital admissions and deaths) potentially suffers from a number
of econometric challenges:
1. Omitted variable bias
If we fail to control for all potential determinants of Covid-19 that are possibly
correlated with pollution, then the estimated coefficients on the pollution measures
could potentially be biased. As will be outlined in Section 4, our strategy is to control
for a wide range of potential explanatory variables and thereby keep any such bias
to a minimum.
2. Measurement error of pollution
This could take 2 forms. First, we may not be accurately capturing long-term pol-
lution exposure within each municipality. While Section 4 outlines how we measure
long-term pollution exposure, we note that any non-systematic measurement error of
this kind is likely to result in a conservative estimate of the association between pollu-
tion and Covid-19 due to possible attenuation bias. Second, there may be systematic
measurement error if the error was in some way related to the pollution-Covid-19
relationship. For instance, if individuals with poor health were more (or less) likely
to live in polluted areas and hence were more (or less) likely to catch or succumb to
the virus. We are unable to identify any potential mechanisms of this nature and
our wide range of explanatory variables should obviously help to control for any such
7effects if they were to exist.
Nevertheless, we do instrument pollution in equation 1 as a means of further reducing
concerns around attenuation bias. Since we are using a non-linear count model we
instrument using a control function approach, which is likely to be more efficient
than a standard instrumental variables model (Wooldridge, 2015). This approach
involves estimating air pollution, our potentially endogenous variable, as a function
of our instruments and other exogenous variables, and then inserting the predicted
errors from this first stage into the second stage as a separate control variable (in
addition to our air pollution variable). A simple test of the statistical significance of
the coefficient on the predicted residuals will inform us whether our pollution variable
was indeed endogenous, a procedure equivalent to a Durbin-Wu-Hausman exogeneity
test.
To be suitable for use as an instrument a variable should be correlated with air pollu-
tion but independent of our Covid-19 variables and hence should only influence them
through the effect of air pollution. We experiment with two potential candidates.
The first is a long lag of air pollution. More specifically, we use annual pollution
concentrations averaged over the period 1995-2000. Second, we use a measure of
the average commuting time for residents in each municipality. This variable draws
upon free-flow road network data from VUGeoPlaza enabling us to calculate average
travel time for every j and k location pair. In order to calculate the commuting times
we obtain the residential and work locations of each worker at the 4 − digit postal
code level (which corresponds to neighbourhood level) by constructing an employer-
employee data set (LEED) based on linking administrative data, Dutch Labour Force
surveys, and Tax Registers.8 We then combine the travel time data with the LEED.
Travel times linked with the LEED allow us to calculate the commuting time of each
worker to their actual place of work. Additionally we calculate the travel time based
on the Euclidean distance between the 4 − digit postal code, assuming an average
speed of 10km/h. We then choose the lowest of the commuting time and the Eu-
clidean distance travel time for each worker and average these across municipalities.9
8
The access to these data sets requires confidentiality agreement with Statistics Netherlands.
9
See Koster & Ozgen (2020) for more details. Generally, for short distances (e.g. less than 5km) it is
8Differences in commuting time between municipalities reflect differences in the dura-
tion and intensity of journeys undertaken and should therefore influence differences in
air pollution concentrations. We see no reason why commuting time should directly
influence Covid-19 cases.
3. Measurement error of Covid-19
A third potential econometric challenge is the existence of possible measurement error
within our Covid-19 data. Non-systematic measurement error would tend to reduce
the precision of our estimates by increasing the standard errors on the estimated coef-
ficients. Alternatively, if the measurement error is related to other omitted variables
that are important for the pollution-Covid-19 relationship, or to its own values, this
could be a cause for concern. Given our rich set of controls and the manner in which
Covid-19 data are collected, this is unlikely to be a problem in our context and hence
we see no reason for the existence of such systematic measurement error. Section 4
discusses the nature of our Covid-19 data in more detail.
4. Spillover effects
The final potential econometric challenge is the possible existence of spillover effects
caused by the virus spreading from one municipality to another. To allow for this
possibility we extend equation 1 to include a spatially lagged dependent variable and
a spatial error term, each using spatial weight matrices with varying distance cut-offs,
as specified in equation 2.
0 0 0 0 0
Ci = θpollutioni + α1 Di + α2 Pi + α3 Ei + α3 Si + α4 Li + ρwi Ci + λwi i + µi (2)
where wi is an inverse distance spatial weight matrix with a 50km cut-off, a 100km
cut-off, or no cut-off at all. All other variables are as previously specified. Note that
the inclusion of a spatially lagged dependent variable into a non-linear count data
model is not straightforward and hence there has yet to be a widely accepted method
of doing so (Glaser, 2017). Equation 2 is therefore a linear model estimated using
slower to use the road network and so the Euclidean travel time is used.
9maximum likelihood.
4 Data and Summary Statistics
4.1 Data
We utilize extensive secondary data combined with administrative micro-data and spa-
tial data from various sources in the Netherlands. Our analysis rests on municipality level
cross-sectional Covid-19 data provided by the National Institute for Public Health and the
Environment (RIVM) and a rich set of controls from 2019 (unless otherwise stated). We
further use high resolution spatial data for air pollution indicators which span the period
2015-19.
The municipal-level data is obtained from the data repository Statline of the Statistics
Netherlands, while some variables are aggregated from the administrative micro-data up
to municipality level. When modelling the relationship between Covid-19 and air pollu-
tion, we control for a wide range of potentially confounding effects which we categorise
as demographic, social and physical proximity, employment and education, spatial, and
health-related. Below we discuss Covid-19 and pollution data and each category of control
variable:
4.1.1 Covid-19
The Covid-19 data used in our analysis is obtained from the RIVM. The Regional Public
Health Service Centres (known as GGD) from all across the country10 provide RIVM with
the Covid-19 data, which is then corrected for any inconsistencies before being released
by RIVM. As stated by RIVM on their website, there is a slight delay between the day
of hospitalization or death and the day on which the the number of cases is reported.11
However, when the data is available RIVM does the necessary adjustment to the data
retrospectively. This incident of slight delay is well-known for all countries. Again similar
to the other countries, the number of Covid-19 cases is likely to be an underestimate as
not everybody in the country is tested. Finally, all Covid-19 variables are coded by the
10
https://www.ggd.amsterdam.nl/coronavirus/.
11
https://www.rivm.nl/coronavirus-covid-19/actueel.
10residence of each individual rather than the address of the test centre or the hospital to
which they are admitted.
4.1.2 Air pollution
Annual concentrations of P M2.5 , N O2 , and SO2 measures are provided by RIVM. These
pollution concentrations are reported at the level of 1x1 km2 grid-cells having been mod-
elled using a wide range of sources and components in the Netherlands and in other Euro-
pean countries. Maps of the spatial distribution of pollution concentrations were calculated
using the Operational Priority Substances dispersion model, which constructs the average
annual concentrations of pollutants stemming from the dispersion, transport, chemical con-
version and dispersion of emissions. The resultant concentrations are calibrated with the
observations from the Dutch Air Quality Monitoring Network G. J. Velders et al. (2020);
Fischer et al. (2020).12
To produce municipality level measures of pollution concentrations we use the median
grid-cell concentrations within each municipality. In order to overcome any potential mea-
surement bias due to annual fluctuations and to capture the long term exposure of residents
within a municipality, we average the annual pollution concentrations data over the 5-year
period 2015-2019.
4.1.3 Demographic and labour market indicators
The demographic controls include the total population as well as the density of popula-
tion in each municipality, expressed as population per square kilometre. The share of the
population over 70 years of age in our regressions accounts for the particular vulnerability
of the elderly to Covid-19, as the stylized facts from around the world indicate higher ca-
sualties among the elderly population. We also include the share of the population under
18 years old, with the omitted category being the working age share of the population.
Finally, several recent studies from a number of countries including the UK and the US13
have indicated that ethnic minorities are disproportionately affected by Covid-19. There
are a number of potential reasons why this might be so. First, Zorlu & Hartog (2012) show
12
For more details on the construction of the pollution concentrations data see G. Velders et al. (2017)
13
See Kirby (2020) and Platt & Warwick (n.d.) https://www.ifs.org.uk/inequality/chapter/are-some-
ethnic-groups-more-vulnerable-to-covid-19-than-others
11that ethnic groups in the Netherlands are more likely to work in manual occupations, which
have been shown to increase the risk of exposure to Covid-19 due to the greater frequency
of face-to-face contact amongst workers (Lewandowski, 2020).14 Second, ethnic minorities
may be more likely to experience social deprivation, live in smaller housing and belong to
larger households. As will be discussed below, we try to control for each of these factors
by taking housing and household size conditions into account. Other risk factors may
include the use of certain cultural practices (e.g., attending places of worship) or having
a disposition for underlying health conditions. To control for any such effects we include
the number of non-western migrants as a share of the population. Non-western migrants
in the Netherlands are defined as the number of immigrants who are born in Africa, Latin
America, Asia (excluding Japan and Indonesia), and Turkey.
We also include the average household income level and the share of employment in
elementary occupations to control for the possibility that groups that are economically
and occupationally more vulnerable could be more affected by Covid-19. Finally, our
specification includes the number of highly educated people as a share of the total educated
population at the municipal level. Highly educated is defined as those who have obtained
a bachelor’s degree or higher.
4.1.4 Social and Physical Proximity
While the majority of countries have disproportionately experienced Covid-19 infections
in large metropolitan areas, the empirical findings so far are unclear about the role played
by population density.15 Epidemiological studies have argued that density per se is not
important but rather it is the type of interaction between people that is important e.g.
weddings, close family events and occasions when individuals come into close proximity
with each other (Hu et al., 2013). One such occasion in the Netherlands is annual the
Carnival which is widely celebrated in the southern regions of the country – North Brabant
and Limburg– as discussed in Section 2. To capture the potential impact of the Carnival we
initially rely on province level fixed effects which capture the effects in North Brabant and
14
https://www.theguardian.com/world/2020/may/11/manual-workers-likelier-to-die-from-covid-19-
than-professionals.
15
See for example,
https://www.centreforcities.org/blog/have-uk-cities-been-hotbeds-of-covid-19-pandemic/
http://jedkolko.com/2020/04/15/where-covid19-death-rates-are-highest/.
12Limburg and all other provinces individually. Furthermore, in sensitivity analysis we test a
Carnival(municipality) measure by creating a dummy variable for the seven most populous
municipalities (Den Bosch, Tilburg, Breda, Bergen op Zoom, Eindhoven, Helmond, Oss)
within the borders of North Brabant and Limburg provinces. Alongside this, in order to
account for the potential spread of the virus from the areas where the Carnival is widely
celebrated, we also test a proximity to Carnival(municipality) measure that is the inverse
distance from the centroid of these seven Carnival municipalities to the centroid of all
municipalities in the Netherlands.
To capture physical interactions between households we include the share of small
housing, defined as the share of total housing that is between 15 and 50 m2 in size, and
the average household size. We also apply the same principle to workplaces. We aggregate
a linked employer-employee data set based on the population of firms and employees in
the Netherlands from 2016 records in order to create an average firm size variable at the
municipal level. This data is obtained from Tax Registers, which include approximately
12 million observations from the universe of employees.
4.1.5 Spatial Controls
Our specification includes several spatial variables which capture the spatial interactions
and spillovers between municipalities. First, we account for the municipalities where Covid-
19 infections may have been influenced by the transmission of the virus beyond Dutch
national borders. The Dutch municipalities bordering Belgium experience substantial daily
cross-border mobility and Covid-19 cases in Belgium have been the highest of all countries
in per capita terms. Dutch municipalities that border Germany also experience intensive
daily crossings, although, in contrast, the Covid-19 cases in Germany had been relatively
low compared to other European countries. In order to capture the potential contagion
through cross-border mobility, we therefore construct 2 dummy variables; municipalities
bordering Germany and municipalities bordering Belgium.
Second, Schiphol airport within the borders of the municipality of Haarlemmermeer (9
km from Amsterdam) is the main international hub of the country and one of the largest
airports in Europe with approximately 72 million passengers a year. We therefore construct
a variable capturing the inverse distance from each municipality centroid to Schiphol airport
13to capture the potential contagion from the high volume of passengers. Lastly, to capture
the effect of the prevailing wind from the North Sea which could mean less physical spread
of the virus and/or reduced spillovers from neighbours, we include a dummy variable for
all of the municipalities on the Dutch coast.
4.1.6 Health-related
Covid-19 statistics in many countries have indicated that those persons with underlying
health conditions are likely to be disproportionately affected by the virus. Our data allows
us to control for these groups at the municipality level. We include in our estimations the
share of smokers and the share of those suffering from obesity in the total local population
in 2019. Moreover, the share of people receiving incapacity benefits is also included to
proxy those who cannot be active in the labour market due to health constraints.
4.2 Descriptive Statistics
We report the descriptive statistics of our data in Table 1. Our Covid-19 data covers
the period since the first Covid-19 case in February until the tail-end of the epidemic in
June. During this period there have been on average 131 Covid-19 cases per municipality
with a maximum number of 2416. In the same period on average 33 hospital admissions
were recorded across all municipalities (maximum 611) while on average 17 people died
(maximum 336).
Our data set includes 355 municipalities based on the 2019 municipal borders in the
Netherlands, where each has an average population of approximately 47000 people, of which
7.4% are non-Western European immigrants. Although the size of the municipalities is
broadly uniform, average population density across municipalities shows a large variation,
ranging from 23 individuals per km2 to 6523 per km2 . On average 3.5% of housing is
classed as small (15-50 m2 ) while the average household size is 2.3 individuals. Finally,
15% of the population is over 70 years old.
The median annual wage is 32722. With 32% of the population holding a university
degree or higher, Netherlands has a relatively highly-educated population. 9% of the pop-
ulation performs elementary occupations which require low-skilled manual-task intensive
work. Median firm size is 556 workers, but becomes larger for peripheral areas, although
14the average firm size in large cities is still around a standard deviation higher than the
median. Although known to be one of the healthiest nations, 19% of the population are
smokers and 14% experience obesity (defined as % of over 19s with BMI¿30 kg/m2). 6%
receive incapacity benefits due to not being able to work.
With regard to our pollution concentrations data which is measured in µg/m3 , we see
that the average value of P M2.5 concentrations in our dataset is 10.491, with a maximum of
12.264. For N O2 concentrations the mean value is 15.757, with a maximum value of 27.413,
and for SO2 concentrations the mean value is 0.797 with the maximum being 3.092. While
the EU air quality standards do not stipulate a safe limit for annual concentrations of SO2 ,
for P M2.5 and N O2 concentrations the limit is 25µg/m3 and 40µg/m3 , respectively. We
can therefore see that 5 year averages of annual concentrations in the Netherlands do not
exceed these limits. However, it is worth noting that the maximum values of the annual
concentrations of our 1 x 1 km2 grid-cells that form the basis of our municipality pollution
data are 23.9µg/m3 and 62.4µg/m3 for P M2.5 and N O2 , respectively. In the case of N O2 ,
this is significantly beyond the EU safe limit, while for P M2.5 it is very close to the safe
limit.
5 Results
Table 2 provides our initial negative binomial estimates of Equation 1 with columns
(1) to (3) providing the estimates of Covid-19 cases for the three pollutants. Columns (4)
to (6) do the same for hospital admissions, and columns (7) to (9) provide estimates of
Covid-19 deaths, again for the three pollutants.
Focusing initially on the estimated pollution coefficients, we see that P M2.5 concentra-
tions have a positive and statistically significant relationship with Covid-19 cases, hospital
admissions and deaths. In the case of N O2 we find a positive and statistically significant
association between Covid-19 cases and deaths, but this is not statistically significant for
hospital admissions. Finally, for SO2 we again discover a positive relationship with our
dependent variables, but this is only statistically significant for Covid-19 deaths. After
calculating marginal effects, we find that a one unit increase in P M2.5 concentrations is as-
sociated with 9.4 more Covid-19 cases, 3.0 more hospital admissions, and 2.3 more deaths.
15Table 1: Summary statistics
Variable Mean Std. Dev. Min. Max. N
PM2.5, 5yr ave 10.491 1.354 6.915 12.264 355
NO2, 5yr ave 15.757 3.861 6.835 27.413 355
SO2, 5yr ave 0.797 0.36 0.21 3.092 355
# of Covid-19 cases 131.459 224.887 0 2416 355
# of deaths 16.865 31.445 0 336 355
# of hospital admissions 33.017 55.129 0 611 355
Carnival(province) 0.262 0.44 0 1 355
Carnival(municip) 0.02 0.139 0 1 355
Proximity to Carnival(municip) 0.000 0.000 0 0.0001 355
Days since 1st case 86.468 3.197 76 89 355
Proximity to Schiphol 86.491 46.151 5.274 194.828 355
Belgian border 0.073 0.261 0 1 355
Coast 0.093 0.291 0 1 355
German border 0.09 0.287 0 1 355
Ave gross income 32.722 3.05 23.5 53.6 355
ln(firm size) 8.9 0.513 3.258 10.06 355
Sh elementary occup 0.089 0.031 0 0.2 358
Sh high educated 0.323 0.095 0.071 0.9 358
Ave household size 2.297 0.178 1.712 3.339 355
Sh of small houses 0.036 0.045 0 0.672 355
Sh of under 18 0.196 0.024 0.126 0.335 355
Sh of 70+ 0.15 0.025 0.063 0.243 355
Obesity 14.389 2.13 9 22 355
Smokers 19.623 2.747 14 31 355
Sh receiving incapacity benefits 0.063 0.017 0.02 0.132 355
Population/km2 877.315 1042.722 23.111 6523.142 355
ln(population) 10.398 0.826 6.842 13.668 355
Non-western mig sh 0.074 0.059 0.014 0.386 355
Ave commuting time 14.066 8.284 4.279 134.432 358
Note: The proximity to carnival measure is calculated as the negative exponential of dis-
tance i.e. (exp(-distance)) and produces numbers that are very small in magnitude, thereby
explaining the large estimated coefficients for this variable.
A one unit increase in N O2 increases Covid-19 cases by 2.2 and deaths by 0.35. To make
these comparable, a one standard deviation increase in P M2.5 and N O2 concentrations
increases Covid-19 cases by 12.7 and 8.6, respectively. The same one standard deviation
in P M2.5 and N O2 concentrations increases Covid-19 deaths by 3.0 and 1.4, respectively.
Table 3 summarises our pollution marginal effects from the estimated results in which
pollution is statistically significant.
Turning to the other explanatory variables, and focusing on those that are generally
16statistically significant, we find that municipalities next to the German border have lower
Covid-19 cases, hospital admissions, and deaths compared to other municipalities. We
also find average household income to have a negative association with all three dependent
variables, while average household size and the share of housing that is small both have
positive relationships with our Covid-19 variables. Relative to the share of the working age
population, the share of those under 18 has a negative association with Covid-19 while,
as expected, the share over 70s has a positive association. Finally, smokers, the share of
non-Western immigrants, and the total population of each municipality are associated with
increased Covid-19 cases, hospital admissions, and deaths.
Since we find the greatest statistical significance and largest marginal effects for P M2.5
concentrations, for reasons of space we focus on this pollutant alone for the remainder of
our analysis.16 Table 4 reports the results of the instrumental variable estimations using
a control function approach, as outlined previously. In Table 4 we see that the estimated
coefficients on P M2.5 concentrations remain positive and statistically significant, with a
similar magnitude to those in Table 2. It is notable that the first stage residuals are not
statistically significant, thus failing to reject the null of exogeneity. One can also see that
an F-test on our instruments in the first stage is highly statistically significant. Table 7 in
the Appendix reports the first stage results, which use commuting time, commuting time
squared, and lagged particulate matter concentrations from 1995-2000 as instruments. In
sensitivity analysis we tested different combinations of these variables including lagged
pollution alone, commuting time alone, and commuting time in linear form only. In each
case the first stage residuals were not statistically significant in the second stage. Our
results therefore provide reassurance that our estimations in Table 2 are not being unduly
influenced by endogeneity.
16
Results for other pollutants are available upon request from the corresponding author.
17Table 2: Main Estimation Results 18
Table 3: Summary of estimated Marginal Effects for pollution
Table 5 reports the results of the spatial econometric estimations based on Equation
2 in which we include a spatially lagged dependent variable and spatial errors using three
different spatial weight matrices (50km, 100km, no cut-off). Table 5 reports the estimated
coefficients on P M2.5 concentrations for the three dependent variables. The three models
with differing weight matrices are presented. For each we also report the spatial error and
spatial lag coefficients, λ and ρ, respectively. Table 5 indicates that P M2.5 concentrations
continue to have a positive and statistically significant relationship with all three of the
Covid-19 dependant variables, for all three of the spatial weight matrices. The introduction
of the spatially lagged dependent variable into Equation 2 changes the interpretation of our
estimated coefficients. P M2.5 concentrations in municipality i continue to affect the condi-
tional mean of Covid-19 in that municipality but now that change in Covid-19 potentially
changes the conditional mean of Covid-19 in other nearby municipalities (depending on our
weight matrix). Furthermore, the change in Covid-19 in those nearby municipalities affects
the conditional mean of Covid-19 in their neighbouring municipalities and so on. P M2.5
concentrations therefore now have a direct effect on Covid-19 in their own municipality
plus an indirect effect on Covid-19 in other municipalities.
If we take the example of our estimate of Covid-19 Cases using a weight matrix with
a 50km cut-off, which reports an estimated coefficient on P M2.5 concentrations of 20.88,
we can calculate the marginal effect to consist of a direct effect of 20.95, an indirect
effect of -5.89, and a total effect of 15.06. This implies that a 1 unit change in P M2.5
19Table 4: Instrumental Variables Results
concentrations results in an increase in Covid-19 cases of 15.06. However, it is notable that
the estimates of both λ and ρ from this model are not statistically significant, implying
that spatial spillovers are not unduly influencing our results, either through the error term
or through the dependent variable. Indeed, if we look at the estimates of λ and ρ from
the other models in Table 5 we see that the majority are not statistically significant. This
is particularly true of the spatial error coefficient λ. So, while we find limited evidence of
statistically significant spatial spillovers, Table 5 at least confirms that our positive and
statistically significant association between Covid-19 and P M2.5 concentrations is robust
to the inclusion of such spillovers.
Finally, Table 6 reports some additional sensitivity estimations. In columns (1) to (3),
rather than relying only on province dummies to control for the effect of the carnival, we
include a municipality-level dummy of the seven most populous municipalities within North
Brabant and Limburg provinces. In addition, we include a measure of the proximity of
each municipality to the centroid of these seven combined municipalities. Table 6 indicates
that these municipality-level carnival variables are not statistically significant, nor do they
have the expected signs. Furthermore, the coefficients on P M2.5 concentrations remain
unaffected by their inclusion in terms of magnitude, sign, or significance.
As mentioned earlier, the effect of urban density on the spread of Covid-19 has been an
ongoing discussion with a number of studies providing mixed and inconclusive evidence.
20Table 5: Sensitivity Checks I - Spatial Results
To explore this further, columns (4) to (6) in Table 6 report estimations in which we
omitted from our analysis the major urban areas of Amsterdam, Rotterdam, Utrecht and
The Hague in case, given their high levels of concentration, they are unduly influencing
our results. The sign, significance and magnitude of our estimated coefficients on pollution
are consistent with our previous results (in columns (1), (4) and (7) from Table 2).
6 Conclusion
This paper has contributed to the nascent literature examining the link between poor
air quality and Covid-19. We examine data for 355 Dutch municipalities to identify the
relationship between concentrations of P M2.5 , N O2 and SO2 and Covid-19 cases, hospital
admissions and deaths. In contrast to much of the previous literature we are able to control
21Table 6: Sensitivity Checks II - Additional Results
for a wide range of potential confounding effects and, by examining Covid-19 data between
February and June 2020, are able to examine almost the full duration of the epidemic
within the Netherlands.
We find compelling evidence of a statistically significant positive relationship between
air pollution and Covid-19 cases, hospital admissions and deaths. This relationship is
particularly evident for concentrations of P M2.5 and to a lesser extent N O2 and persists
after controlling for explanatory variables capturing income, demography, social and phys-
ical proximity, employment/education, health and spatial factors. The relationship with-
22stands a number of sensitivity and robustness exercises including instrumenting pollution
to mitigate potential endogeneity and modelling spatial spillovers using spatial econometric
techniques.
Our results indicate that an increase in P M2.5 concentrations of 1µg/m3 is associated
with an increase in Covid-19 cases of between 9.4 and 15.1, depending on our model. The
same increase in P M2.5 is associated with an increase in Covid-19 hospital admissions of
between 2.9 and 4.4, and an increase in Covid-19 deaths of between 2.2 and 3.6. The only
comparable study to our own is provided by Wu et al. (2020) who examine Covid-19 deaths
in the US. They find that a 1 µg/m3 increase in P M2.5 is associated with an 8% increase
in the Covid-19 death rate. Since the mean number of deaths in our sample is 16.86, our
estimated increases of between 2.2 and 3.6 are equivalent to increases of between 13.0%
and 21.4%, which are clearly larger in magnitude than those of Wu et al. (2020).
Two clear policy implications arise from our analysis. First, the impact of poor air
quality on Covid-19 morbidity and mortality represents a considerable and unexpected
additional cost from air pollution. Our results would therefore suggest a clear need to
regulate air pollution more stringently, even in a relatively well-regulated nation such as the
Netherlands. Second, our findings should prove useful to public health officials by signaling
where subsequent waves of Covid-19, or indeed future pandemics, might hit hardest. This
could be particularly important in countries where pollution sources don’t correlate well
with major metropolitan areas. If, for instance, livestock production is a major source
of emissions, as appears to be the case in some parts of the Netherlands, then advance
warning may prove particularly beneficial for rural areas where healthcare infrastructure
and coordination may be less developed.
We believe the statistical relationships we have observed between P M2.5 concentrations
and Covid-19 data are robust. Furthermore, given the established literature linking poor
air quality with respiratory disease, they are also plausible. Nevertheless, a degree of
caution is needed. Until detailed individual-level data is available providing information
on Covid-19 and a wide range of other individual characteristics the statistical evidence
will remain suggestive, perhaps strongly suggestive, but not conclusive.
23References
Abbey, D., Nishino, N., McDonnell, W., Burchetter, R., Knutsen, S., Beeson, W., & Yang,
J. (1999). Long-term inhalable particles and other air pollutants related to mortality
in nonsmokers. American Journal of Respiratory and Critical Care Medicine, 159 (2),
373-382.
Andree, B. P. J. (2020). Incidence of covid-19 and connections with air pollution exposure
evidence from the netherlands (Working Paper No. 9221). World Bank Group.
Cicala, S., Holland, S. P., Mansur, E. T., Muller, N. Z., & Yates, A. J. (2020, May).
Expected health effects of reduced air pollution from covid-19 social distancing (Working
Paper No. 27135). National Bureau of Economic Research.
Cole, M., Elliott, R., & Liu, B. (2020). The impact of the wuhan covid-19 lockdown on
air pollution and health: A machine learning and augmented synthetic control approach
(Working Paper No. 20-10). University of Birmingham, Department of Economics.
Conticini, E., Frediani, B., & Caro, D. (2020). Can atmospheric pollution be considered a
co-factor in extremely high level of SARS-CoV-2 lethality in Northern Italy? Environ-
mental Pollution, 261 .
Faustini, A., Rapp, R., & Forastiere, F. (2014). Nitrogen dioxide and mortality: review
and meta-analysis of long-term studies. European Respiratory Journal , 44 (3), 744–753.
Fischer, P. H., Marra, M., Ameling, C. B., Velders, G. J., Hoogerbrugge, R., de Vries, W.,
. . . Houthuijs, D. (2020). Particulate air pollution from different sources and mortality
in 7.5 million adults — the dutch environmental longitudinal study (duels). Science of
the Total Environment, 75 .
Glaser, S. (2017). A review of spatial econometric models for count data (Discussion Papers
in Business, Economics and Social Sciences No. 19-2017). Universität Hohenheim.
Hoffman, J. (2004). General Linear Models: An Applied Approach. Pearson, Boston.
Hu, H., Nigmatulina, K., & Eckhoff, P. (2013). The scaling of contact rates with population
density for the infectious disease models. Mathematical Biosciences, 244 (2), 125 - 134.
24Katanoda, K., Sobue, T., Satoh, H., Tajima, K., Suzuki, T., Nakatsuka, H., . . . Tominaga,
S. (2011). An association between long-term exposure to ambient air pollution and
mortality from lung cancer and respiratory diseases in japan. Journal of Epidemiology,
21 (2), 132-143.
Kirby, T. (2020). Evidence mounts on the disproportionate effect of COVID-19 on ethnic
minorities. The Lancet Respiratory Medicine, 8 (20), 547 - 548.
Koster, H., & Ozgen, C. (2020). Cities and Tasks (forthcoming Discussion Paper). IZA
Institute of Labor Economics.
Lewandowski, P. (2020). Occupational exposure to contagion and the spread of covid-19 in
europe (Discussion paper No. 13227). IZA Institute of Labor Economics.
Long, S. (1997). Regression Models for Categorical and Limited Dependent Variables. Sage,
Thousand Oaks, CA.
Ming Han, J., Yun, Y., Li, G., & Sang, N. (2015). Acute nitrogen dioxide (no2) exposure
enhances airway inflammation via modulating th1/th2 differentiation and activating jak-
stat pathway. Chemosphere, 120 , 722 - 728.
Ogen, Y. (2020). Assessing nitrogen dioxide (NO2) levels as a contributing factor to
coronavirus (COVID-19) fatality. Science of the Total Environment, 726 .
Platt, L., & Warwick, R. (n.d.). Are some ethnic groups more vulnerable to COVID-19
than others? (Report). The Institute for Fiscal Studies.
Travaglio, M., Popovic, R., Yu, Y., Leal, N., & Martins, L. M. (2020). Links between air
pollution and COVID-19 in England. medRxiv, June.
Velders, G., Aben, J., Geilenkirchen, G., Den Hollander, H., Nguyen, V. d. S. E., L.,
De Vries, W., & Wichink Kruit, R. (2017). Large-scale air quality concentration and
deposition maps in the netherlands (RIVM Report No. 2017-0117). National Institute
for Public Health and the Environment, Bilthoven, the Netherlands.
Velders, G. J., Maas, R. J., Geilenkirchen, G. P., de Leeuw, F. A., Ligterink, N. E.,
Ruyssenaars, P., . . . Wesseling, J. (2020). Effects of european emission reductions on air
25quality in the netherlands and the associated health effects. Atmospheric Environment,
221 .
Weerdt, A. D., Janssen, B. G., Cox, B., Bijnens, E. M., Vanpoucke, C., Lefebvre, W., . . .
Jorens, P. G. (2020). Pre-admission air pollution exposure prolongs the duration of
ventilation in intensive care patients. Intensive Care Medicine, (in press).
Wooldridge, J. M. (2015). Control function methods in applied econometrics. Journal of
Human Resources, 50 , 420-445.
Wu, X., Nethery, R. C., Sabath, B. M., Braun, D., & Dominici, F. (2020). Exposure to air
pollution and COVID-19 mortality in the United States: A nationwide cross-sectional
study. medRxiv, April .
Zorlu, A., & Hartog, J. (2012). Employment assimilation of immigrants in the netherlands:
Dip and catchup by source country. International Journal of Population Research, 75 .
26A Appendix
Figure 2: Covid-19 hospital admissions and deaths per capita
Figure 3: Ammonia (N H3 ) Concentrations (2019)
27Table 7: First Stage Results
28You can also read
